[Model] Add support for Arcee Foundational Model (#8154)

docs/supported_models/generative_models.md

@@ -45,6 +45,7 @@ in the GitHub search bar.
 | **SmolLM** (135M–1.7B)            | `HuggingFaceTB/SmolLM-1.7B`                      | Hugging Face’s ultra-small LLM series (135M–1.7B params) offering surprisingly strong results, enabling advanced AI on mobile/edge devices. |
 | **GLM-4** (Multilingual 9B)        | `ZhipuAI/glm-4-9b-chat`                          | Zhipu’s GLM-4 series (up to 9B parameters) – open multilingual models with support for 1M-token context and even a 5.6B multimodal variant (Phi-4V). |
 | **MiMo** (7B series)               | `XiaomiMiMo/MiMo-7B-RL`                         | Xiaomi's reasoning-optimized model series, leverages Multiple-Token Prediction for faster inference. |
+| **Arcee AFM-4.5B**               | `arcee-ai/AFM-4.5B-Base`                         | Arcee's foundational model series for real world reliability and edge deployments. |
 | **Persimmon** (8B)               | `adept/persimmon-8b-chat`                         | Adept’s open 8B model with a 16K context window and fast inference; trained for broad usability and licensed under Apache 2.0. |
 | **Granite 3.0, 3.1** (IBM)               | `ibm-granite/granite-3.1-8b-instruct`                          | IBM's open dense foundation models optimized for reasoning, code, and business AI use cases. Integrated with Red Hat and watsonx systems. |
-| **Granite 3.0 MoE** (IBM)               | `ibm-granite/granite-3.0-3b-a800m-instruct`                          | IBM’s Mixture-of-Experts models offering strong performance with cost-efficiency. MoE expert routing designed for enterprise deployment at scale. |
+| **Granite 3.0 MoE** (IBM)               | `ibm-granite/granite-3.0-3b-a800m-instruct`                          | IBM’s Mixture-of-Experts models offering strong performance with cost-efficiency. MoE expert routing designed for enterprise deployment at scale. |

python/sglang/srt/models/arcee.py

+# Copyright 2023-2024 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Inference-only Arcee Foundational Model (AFM) compatible with HuggingFace weights."""
+
+import logging
+from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+from transformers import LlamaConfig
+
+from sglang.srt.distributed import (
+    get_pp_group,
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+)
+from sglang.srt.layers.activation import get_act_fn
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    ColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
+from sglang.srt.layers.pooler import Pooler, PoolingType
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.rotary_embedding import get_rope
+from sglang.srt.layers.utils import PPMissingLayer, get_layer_id
+from sglang.srt.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from sglang.srt.managers.schedule_batch import global_server_args_dict
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
+from sglang.srt.model_loader.weight_utils import (
+    default_weight_loader,
+    kv_cache_scales_loader,
+    maybe_remap_kv_scale_name,
+)
+from sglang.srt.utils import add_prefix, make_layers
+
+logger = logging.getLogger(__name__)
+
+
+class ArceeMLP(nn.Module):
+    """
+    MLP block for the Arcee model, using a ReLU-squared activation function.
+    This differs from the Llama SwiGLU activation.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        reduce_results: bool = True,
+    ) -> None:
+        super().__init__()
+        # Arcee uses a single up-projection, not a merged gate/up projection.
+        self.up_proj = ColumnParallelLinear(
+            hidden_size,
+            intermediate_size,
+            bias=False,
+            quant_config=quant_config,
+            prefix=add_prefix("up_proj", prefix),
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            prefix=add_prefix("down_proj", prefix),
+            reduce_results=reduce_results,
+        )
+        if hidden_act != "relu2":
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. "
+                "Arcee model in SGLang only supports 'relu2'."
+            )
+        # The activation function is relu(x)^2
+        self.act_fn = get_act_fn("relu2")
+
+    def forward(self, x, forward_batch=None):
+        x, _ = self.up_proj(x)
+        x = self.act_fn(x)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class ArceeAttention(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        rope_is_neox_style: bool = True,
+        max_position_embeddings: int = 8192,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        bias: bool = False,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        self.head_dim = getattr(config, "head_dim", None)
+        if self.head_dim is None:
+            self.head_dim = self.hidden_size // self.total_num_heads
+        self.partial_rotary_factor = getattr(config, "partial_rotary_factor", 1)
+        self.rotary_dim = int(self.partial_rotary_factor * self.head_dim)
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=add_prefix("qkv_proj", prefix),
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=add_prefix("o_proj", prefix),
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.rotary_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+            is_neox_style=rope_is_neox_style,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+            quant_config=quant_config,
+            prefix=add_prefix("attn", prefix),
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, forward_batch)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class ArceeDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling is not None and getattr(
+            config, "original_max_position_embeddings", None
+        ):
+            rope_scaling["original_max_position_embeddings"] = (
+                config.original_max_position_embeddings
+            )
+        rope_is_neox_style = getattr(config, "rope_is_neox_style", True)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
+        attention_bias = getattr(config, "attention_bias", False) or getattr(
+            config, "bias", False
+        )
+        self.self_attn = ArceeAttention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            rope_is_neox_style=rope_is_neox_style,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            prefix=add_prefix("self_attn", prefix),
+            bias=attention_bias,
+        )
+        self.mlp = ArceeMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            prefix=add_prefix("mlp", prefix),
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            forward_batch=forward_batch,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class ArceeModel(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.pp_group = get_pp_group()
+        if self.pp_group.is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                prefix=add_prefix("embed_tokens", prefix),
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        self.layers, self.start_layer, self.end_layer = make_layers(
+            config.num_hidden_layers,
+            lambda idx, prefix: ArceeDecoderLayer(
+                config=config, quant_config=quant_config, layer_id=idx, prefix=prefix
+            ),
+            pp_rank=self.pp_group.rank_in_group,
+            pp_size=self.pp_group.world_size,
+            prefix="model.layers",
+        )
+
+        if self.pp_group.is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer(return_tuple=True)
+        self.layers_to_capture = []
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, List[torch.Tensor]], PPProxyTensors]:
+        if self.pp_group.is_first_rank:
+            if input_embeds is None:
+                hidden_states = self.embed_tokens(input_ids)
+            else:
+                hidden_states = input_embeds
+            residual = None
+        else:
+            assert pp_proxy_tensors is not None
+            hidden_states = pp_proxy_tensors["hidden_states"]
+            residual = pp_proxy_tensors["residual"]
+
+        aux_hidden_states = []
+        for i in range(self.start_layer, self.end_layer):
+            if i in self.layers_to_capture:
+                aux_hidden_states.append(hidden_states + residual)
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                forward_batch,
+                residual,
+            )
+
+        if not self.pp_group.is_last_rank:
+            return PPProxyTensors(
+                {
+                    "hidden_states": hidden_states,
+                    "residual": residual,
+                }
+            )
+        else:
+            hidden_states, _ = self.norm(hidden_states, residual)
+
+        if len(aux_hidden_states) == 0:
+            return hidden_states
+
+        return hidden_states, aux_hidden_states
+
+    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
+        tp_size = get_tensor_model_parallel_world_size()
+        tp_rank = get_tensor_model_parallel_rank()
+        for layer_idx, scaling_factor in kv_cache_scales_loader(
+            quantization_param_path,
+            tp_rank,
+            tp_size,
+            self.config.num_hidden_layers,
+            self.config.__class__.model_type,
+        ):
+            if not isinstance(self.layers[layer_idx], nn.Identity):
+                layer_self_attn = self.layers[layer_idx].self_attn
+
+            if hasattr(layer_self_attn.attn, "k_scale"):
+                layer_self_attn.attn.k_scale = scaling_factor
+                layer_self_attn.attn.v_scale = scaling_factor
+            else:
+                raise RuntimeError(
+                    "Self attention has no KV cache scaling factor attribute!"
+                )
+
+
+class ArceeForCausalLM(nn.Module):
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        # Note: gate_proj is removed compared to Llama
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    # in TP, these weights are partitioned along the column dimension (dim=-1)
+    column_parallel_weights_modules = [".down_proj.", ".o_proj."]
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        # Note: gate_proj and up_proj are removed as they are not stacked in ArceeMLP
+        ".q_proj": (".qkv_proj", 0),
+        ".k_proj": (".qkv_proj", 1),
+        ".v_proj": (".qkv_proj", 2),
+    }
+
+    def __init__(
+        self,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.pp_group = get_pp_group()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = self._init_model(config, quant_config, add_prefix("model", prefix))
+        # Arcee does not tie word embeddings
+        self.lm_head = ParallelLMHead(
+            config.vocab_size,
+            config.hidden_size,
+            quant_config=quant_config,
+            prefix=add_prefix("lm_head", prefix),
+            use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
+        )
+        self.logits_processor = LogitsProcessor(config)
+        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
+        # Parameters that are stacked in a single tensor in this model
+        self.stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            (".qkv_proj", ".q_proj", "q"),
+            (".qkv_proj", ".k_proj", "k"),
+            (".qkv_proj", ".v_proj", "v"),
+        ]
+        self.capture_aux_hidden_states = False
+
+    def _init_model(
+        self,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        return ArceeModel(config, quant_config=quant_config, prefix=prefix)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+        get_embedding: bool = False,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> LogitsProcessorOutput:
+        hidden_states = self.model(
+            input_ids,
+            positions,
+            forward_batch,
+            input_embeds,
+            pp_proxy_tensors=pp_proxy_tensors,
+        )
+
+        aux_hidden_states = None
+        if self.capture_aux_hidden_states:
+            hidden_states, aux_hidden_states = hidden_states
+
+        if self.pp_group.is_last_rank:
+            if not get_embedding:
+                return self.logits_processor(
+                    input_ids,
+                    hidden_states,
+                    self.lm_head,
+                    forward_batch,
+                    aux_hidden_states,
+                )
+            else:
+                return self.pooler(hidden_states, forward_batch)
+        else:
+            return hidden_states
+
+    @property
+    def start_layer(self):
+        return self.model.start_layer
+
+    @property
+    def end_layer(self):
+        return self.model.end_layer
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.model.embed_tokens
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        params_dict = dict(self.named_parameters())
+
+        for name, loaded_weight in weights:
+            layer_id = get_layer_id(name)
+            if (
+                layer_id is not None
+                and hasattr(self.model, "start_layer")
+                and (
+                    layer_id < self.model.start_layer
+                    or layer_id >= self.model.end_layer
+                )
+            ):
+                continue
+            if "rotary_emb.inv_freq" in name or "projector" in name:
+                continue
+            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
+                continue
+
+            # Handle FP8 kv-scale remapping
+            if "scale" in name:
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+
+            is_stacked = False
+            for param_name, weight_name, shard_id in self.stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+
+                name = name.replace(weight_name, param_name)
+                if name not in params_dict:
+                    continue
+
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                is_stacked = True
+                break
+
+            if not is_stacked:
+                if name in params_dict:
+                    param = params_dict[name]
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
+                    weight_loader(param, loaded_weight)
+                else:
+                    logger.warning(f"Parameter {name} not found in model.")
+
+    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
+        self.model.load_kv_cache_scales(quantization_param_path)
+
+
+EntryClass = [ArceeForCausalLM]